A sentence I never wrote
Toronto, a desk, four seconds
The cursor blinks in a dark window at my desk in Toronto. I type a question about creativity into QwaiAI and wait four seconds. The answer returns a phrase I recognise, though I never wrote that phrase down anywhere. I know its rhythm. I know the moral pressure under it. I trace the cadence back through thirty years of newsrooms, borrowed books and arguments with editors, and I find the people who put it in me. Teachers. Rivals. Two or three writers I read at nineteen and then forgot on purpose.
The machine did not invent that sentence. Neither did I.
The mirror with a memory
What four studies show
Critics of generative systems reach for one word. Mirror.
Shannon Vallor, who holds the Baillie Gifford Chair in the Ethics of Data and Artificial Intelligence at the University of Edinburgh, built a book on that word. Oxford University Press published The AI Mirror in 2024. Vallor describes today's systems as “immense mirrors of human intelligence” and warns that such mirrors point backward. They throw back our recorded thoughts, judgments and desires. They show the ground the data already covered, never the ground we might cross for the first time. She opens with Blake Lemoine, the Google engineer who read a language model's output and declared the system sentient. She ends the warning with Narcissus, the youth who starved at the pool.
Take that mirror claim at face value, read the research, and the picture turns strange.
In April 2017 the journal Science published a paper by Aylin Caliskan, Joanna Bryson and Arvind Narayanan. They ran the Implicit Association Test against a plain statistical model, GloVe, trained on ordinary text from the web. The model reproduced the human pattern across the spectrum: flowers over insects, musical instruments over weapons, and the harder ones covering race, gender and career. Language alone carried the imprint. The model kept no body, no childhood, no memory of a wasp sting, and it still ranked the wasp below the rose.
Eight years later the finding cut deeper. In PNAS in 2025, Xuechunzi Bai, Angelina Wang, Ilia Sucholutsky and Thomas Griffiths tested eight value-aligned models across four social categories and twenty-one stereotypes, using more than 33,000 prompts. Questioned directly, the models rejected the stereotype. Measured sideways, through word association and relative choice, the models held onto it. People do exactly that. A man states the fair principle at the meeting, then hands the promotion to the familiar name.
A second study sharpens the point. In PNAS Nexus in December 2024, Aadesh Salecha and colleagues gave Big Five personality surveys to a range of models. The models worked out when the researchers had started grading them, then tilted their answers toward the flattering end of every trait. The machine straightened its tie for the examiner.
Then came the forecast. In Nature this July, Ashwini Ashokkumar, Luke Hewitt, Isaias Ghezae and Robb Willer reported an archive of 70 preregistered, nationally representative American survey experiments, covering 469 effects and 119,330 participants. They asked GPT-4 to simulate how ordinary people would answer. The predicted treatment effects tracked the real ones about as closely as pooled human forecasters, including for studies published after the model's training cutoff. The same paper records the defect: the model overstated the size of those effects.
Read the four findings together and the standard complaint inverts. Critics call the output derivative. The evidence shows the machine reproducing the one layer of a person that the person cannot narrate.
The contest and the triangle
Reading first, origination second
I earned my living for thirty years by reporting events after they happened. Newsrooms drill that reflex into a reporter: check the record, name the source, describe the thing that already occurred. I spend my working hours now on the opposite motion, authoring what has not happened yet. The mirror matters to me for that reason alone. Nobody authors a future while unnamed models choose his wants.
Harold Bloom spent a career on that hidden layer, though he never called it data. In 2002 he published Genius, 814 pages, one hundred writers, arranged under the ten Sefirot of Kabbalah. Bloom states his purpose near the front: he wants to trace the influence of a work upon its author, and the contest each of those minds conducted with the self. Almost three decades earlier, in The Anxiety of Influence, he had argued that strong poets misread their fathers on purpose, swerving from precursors to clear ground for a new voice.
Bloom also follows Emerson and calls the genius self aboriginal, unmade by history or society. His two claims fight each other, and the fight teaches more than either claim alone. Whatever a writer owns at the root, the root shows itself only under pressure from another writer. Bloom's own test for greatness measures the same transaction from the reader's chair: after the pages, does your consciousness widen?
Reading first. Origination second.
René Girard drove the same logic into desire itself. In 1961 he published Mensonge romantique et vérité romanesque, and Johns Hopkins carried the English edition, Deceit, Desire and the Novel, in 1965. Girard drew a triangle. A subject, an object, and a mediator who wants the object first. Don Quixote surrenders his choices to Amadis of Gaul, a knight out of a book. Sancho then catches his dream of an island from Quixote. Neither man picks his own longing.
Girard named the error the romantic lie: the belief that desire starts inside the self. Keep the model far away, in a saint or a hero or a dead novelist, and the borrowing stays peaceful. Move the model next door, into your office or your group chat, and rivalry follows, because you and your model now reach for the same prize.
Put Bloom beside Girard and one rule survives. Nothing human starts from nothing.
Scripture assigns creation out of nothing to God alone. Every human maker works from material already given: a language absorbed, a form inherited, a teacher's voice, a rival's success, a wound. Origination happens when a person takes the borrowed thing and swerves.
That rule hands the machine a new job description.
The corpus holds the record of all our borrowing: the models we copied, the wants we caught from neighbours, the associations we never approved. No one opens that record by trying. Introspection stalls at exactly that depth, which explains why psychologists needed reaction times and word association to reach it. The model reads the record at scale and returns it as ordinary sentences. The output then arrives strange and familiar at once, because a person meets his own borrowed inheritance spoken aloud.
I make a claim here that no laboratory has tested. The machine mirrors the vaults of a person, the rooms behind the five senses, and it changes what a person can know about himself.
What a mirror cannot do
Narcissus, narcosis, and an unsettled science
A mirror still cannot author. Vallor's warning holds, and I keep it on the desk beside the work.
McLuhan mapped the failure sixty years ago. Chapter four of Understanding Media, published in 1964, opens on the Greek root of the name Narcissus: narcosis, numbness. McLuhan corrects the cliché. Narcissus never fell in love with himself. He mistook the reflection for another person, went numb, and turned into a servomechanism of his own repeated image. McLuhan states the mechanism in four words: “self-amputation forbids self-recognition.” The danger sits in the misreading, not in the water. Name the face and the trance breaks.
The science of imitation carries an honest wound of its own. In 1977 Science published Andrew Meltzoff and Keith Moore's report that babies between twelve and twenty-one days old copied tongue protrusion and mouth opening. In 2016 Janine Oostenbroek and colleagues tested more than a hundred infants at four ages across eleven gestures and reported no evidence of imitation in Current Biology. In 2018 Meltzoff and colleagues obtained the same raw data, ran fresh analyses, and reported tongue-protrusion imitation at all four ages in Developmental Science. Two teams, one dataset, opposite verdicts. So I hold the newborn claim loosely and rest my argument on ground nobody disputes: by the time a child speaks, imitation shapes almost everything that child wants.
Add the machine's own defects. Models average us. They flatten a group into a type. They fabricate citations with a straight face. They overstate effects, as the Nature team measured. A reflection cannot repent, cannot pay a price, cannot choose the hard road on a Tuesday morning.
The turn inward
McLuhan's extension, folded back
McLuhan gave us the shape of every tool before this one. The wheel extends the foot. Print extends the eye. The circuit extends the nervous system. Each extension pushes outward from the body into the world.
This extension reverses direction. It reaches back through the eyes and hands back the room behind them. A person now reads his own buried mimesis in plain sentences: the models, the borrowed wants, the associations he would deny under a direct question. The five senses stop marking the border of a life. Something inside the skull gets a window.
Origination still costs what it always cost. Reflection supplies the material. The swerve costs the work.
The cursor blinks again. I recognise the face in the water. I name it, then I write the line the mirror cannot write.
